Speech Signal Processing Based On Morphological Approaches

The digital speech processing, as the very core technology of the information field, is a broad intercrossing subject. It is indicated that the speech signal is actually a complicated non-linear and non-stationary signal. The non-linear process methods developing and improving rapidly in resent years has compensated the shortcomings of the traditional linear process methods, as they were used on the speech signal processing domain step by step.Digital morphological filter is one of the most important and rapidly developing non-linear filter, which had successfully applied in many research fields, such as image analysis and processing, biomedical signal processing, etc. The application of the morphological filter on speech signal processing had already been explored, but they mostly remained in image processing stage, such as the spectrogram processing, transforming one-dimensional signal into two-dimensional image signal then morphological filtering, etc. They mostly have very large computing complication.The basic frequency of speech signal is an important parameter of the speech signal processing domain. There are many algorithms proposed and applied on pitch detection at present, but they still have many limitations. In this paper, we focus on the studying of the pitch detection methods, using the morphological filter to the one-dimensional speech waveform directly. Two pitch detection method based on morphological filter are proposed. The one is to combine the morphological non-linear filter with the traditional linear filter, using line and half-sin function as structuring element of the morphological filter. And then get the pitch track by auto-correlation function algorithm. The other is to use flat structuring element with long length, getting the picks or valleys of the waveform directly. Furthermore a pitch track smooth method based on morphological filter was proposed. All the algorithms were tested and simulated; the data was analyzed and compared with some traditional methods. The results show that the morphological speech processing is feasible and superior to the traditional method in some aspect. For the briefness in algorithm, the directness in results and the quickness in calculating, morphological filtering will be a preferable approach to non-linear speech signal processing.